| Home > Publications database > AIcXR+: Spiking Neural Network (SNN) Semantic Compression for 3-D Point Clouds |
| Journal Article | DKFZ-2026-02608 |
; ; ; ; ; ; ;
2026
IEEE
New York, NY
Abstract: Existing Artificial Intelligence (AI) approaches for compressing three-dimensional (3-D) pointclouds for real-time immersion over communication networks are commonly based on energy-intensive AIarchitectures, e.g., transformer and autoencoder architectures. We design and evaluate the first 3-D pointcloud compression based on the energy-efficient Spiking Neural Network (SNN) architecture. We conductthis evaluation within our framework of AI-based semantic compression for networked eXtended Reality(AIcXR+). We first design a goal-oriented autoencoder, wherein an Artificial Neural Network (ANN)encodes a 3-D point cloud into a compact Vector Code (VC) at the sender, and a Convolutional NeuralNetwork (CNN) decodes the VC into volumetric pixels (voxels) for rendering at the receiver. We jointlytrain the ANN encoder and CNN decoder to map 3-D point cloud frames into VCs that compactly representthe volumetric semantics of the 3-D point clouds. Then, we convert this trained ANN/CNN into an SNN,bypassing the standard backpropagation training of SNNs. For robotic surgery datasets, we evaluate the SNNrelative to the ANN/CNN as well as a (non-AI-based) semantic color-coding approach (cXR+). We findthat both SNN- and ANN-AIcXR+ achieve approximately the same compression ratios as cXR+, whileachieving substantially shorter computing (encoding + decoding) latencies than cXR+; thus, AIcXR+substantially reduces the aggregate (computing + communication network transmission) latency comparedto the non-AI-based cXR+. Importantly, we find that the SNN reduces the energy consumption by nearlyan order of magnitude compared to the ANN while achieving essentially the same reconstruction quality.
|
The record appears in these collections: |